High-risk § 3.1.2.2

Article 26(10) AI Act consists of seven subparagraphs that concern three specific situations²¹ . Subparagraphs 1 and 2 contain a specific authorisation requirement for deployers of post-remote biometric identification (post-RBI) systems used for the targeted search of a person suspected or convicted of a crime in the framework of an investigation. Subparagraphs 3 to 6 contain general rules and procedural requirements directed at deployers of any high-risk post-RBI system used for any law enforcement purpose, including for the remote identification and localisation of offenders, the search for missing persons , and the prevention of crimes. Sub paragraph 7 authorises Member States to introduce more restrictive laws on the use of post-RBI.

The use of a high-risk AI system for post-RBI for the purpose of law enforcement , in the framework of an investigation for the targeted search of a person suspected or convicted of having committed a criminal offence, requires the deployer ( generally, a law enforcement authority) to ask for an authorisation by a judicial authority or by an administrative authority whose decision is binding and subject to judicial review. The authorisation requirement does not apply to all RBI systems classified under Point 1(a) of Annex III AI Act. As mentioned above, the scope of its application is narrowed down to specific situations in the framework of an investigation for the targeted search for law enforcement purposes.

The term ‘targeted ’ in ‘targetedsearch’ relates to the question whether the law enforcement authority has reasonable cause to suspect a connection between a suspect and a crime and is therefore searching for that suspect. Once this has been established, the law enforcement authority may use the most efficient RBI search method in view of the concrete situation. It may run , for example, probe or gallery searches, depending on the concrete circumstances.The term ‘search’ in ‘targeted search’ relates to the objective of identifying the location in time and place of the targeted individual.

A search for a person may be targeted whether the identity of the person is known or unknown. A search can be considered targeted where there is a reason to search for a specific person. The suspect must be directly linked to a criminal offence as a suspect or convict. Example of targeted searches:

• Following an armed robbery, investigators obtain CCTV images and identify a known individual, already present in police records, as the prime suspect. Authorities run a post-RBI query using the suspect’s facial template to determine whether the person appears in footage from nearby transport hubs at that moment. The search is targeted as the search is narrowly focused on the one individual, and there is a direct link to a concrete criminal offence.
• A witness to a kidnapping provides investigators with video showing the suspect’s face. Police extract a biometric template from the footage and run a post-RBI search to determine whether that same individual appears in footage captured by other cameras along the suspected escape route. On untargeted search see paras. 159-160 below.

The initial identification of a potential suspect, based on objective and verifiable facts directly linked to an offence, is excluded from the authori sation requirement under Article 26(10) AI Act. The AI Act does not define when a person is considered a suspect , but in the context of Article 26(10) AI Act this should be understood to mean persons who are reasonably believed by law enforcement authorities to have committed a criminal offence. The suspicion of the law enforcement authorities must be based on objective and verifiable facts. Initial identification relates to the identification of one or several potential suspects against a reference database to find out their identity or to establish the existence of a matching biometric profile in a reference database, without necessarily establishing the suspect’s identity. The question is: Who is the person (that is a potential suspect?). Examples are the comparison of a suspect’s biometric data against biometric templates of known persons, for instance from police files or EU information systems. The objective can be to find out the likely identity of the suspect (e.g. score 8/10 that thi s is Person A) or to fi nd out whether the suspect has already previously committed crimes or was a suspect (this is a person whose biometric traces were already found previously at a crime scene; this is a person whose biometric profile is already on a reference database). In other words, initial identification is not necessarily aimed at establishing the civil identity of a person. In the case of initial identification, the system needs to comply with the requirements for high-risk AI systems, provided that the AI system constitutes a remote biometric identification system and therefore classifies as high-risk (see Section 3.1.2.1. above). No specific authorisation is needed for the initial identification of a potential suspect based on objective and verifiable facts directly linked to the offence. Examples of initial identification:

• After a stabbing, investigators obtain clear CCTV footage showing the perpetrator’s face. Police generate a biometric template and compare it against national police and EU databases to determine whether a matching biometric profile exists and to identify the likely person.
• Following a burglary, police extract a facial image of the perpetrators from CCTV and compare it against biometric traces previously collected from other crime scenes to determine whether the same individual may be involved in multiple offences.
• After a terrorist incident, investigators obtain video showing one perpetrator but cannot yet confirm his or her identity. Post-RBI is used against police databases to produce ranked candidate matches (e.g. ‘likely person A’) to which other pieces of evidence must be corroborated to identify the person.
• Searches carried out within reference databases, including police records, and public or private video surveillance camera systems, by comparing a given biometric sample of an individual, such as (a) pictures of the face, of tattoos, or any other identifiable body part, (b) gait, (c) voice, collected from the following sources and in the following situations (i) from video surveillance cameras , (ii) from human remains, whereabouts or traces notably from unknown persons, (iii) from unconscious persons, (iv) from social media, (v) from seized material , (vi) when taken in custody with coercion , (vii) from legal interception for the first identification in the context of an active search of a person. Examples of initial identification include:
• Airport authorities searching their video-surveillance camera system for the presence (the location) of a given person on the basis of pictures collected from social media, upon the request of national law enforcement authorities searching for that individual. However, upon a match in the system, further searches in the system for the individual and for the same purpose of searching this suspect are not considered to be initial identification.
• Member State B law enforcement authorities consulted by Member State A law enforcement authorities as part of a police cooperation information exchange for the first time in the context of an active search of a person suspected of having committed a crime and searching for the identity of a person in the criminal database is initial identification.
• A search carried out in another Member State ’s national criminal database through the Prüm II framework with latent fingerprints or latent palm prints, or with facial images collected on a crime scene, with the aim of identifying the suspect subject of an active search.
• Europol searches in Member States’ database under the Prüm II framework with fingerprints or facial images received from third-country authorities for establishing the identity of the individual actively tracked until a match occurs .
• Europol, while supporting Member States active search for a given individual suspect of having committed a crime under Europol’s mandate investigation comprising several ATM attacks, carries out comparisons in its own data on the basis of footage from the local surveillance camera. Such processing for identification purposes are considered initial so long as the person is not identified. Further identifications with the same sample into Europol data and with the same purpose of searching that suspect, and potentially revealing the presence of the person in a new setting, cannot be considered initial.

The authorisation for a targeted search is required when the investigation has been narrowed down from a broader number of persons of interest to a limited number of suspects and this is reflected in some formal status of an open investigation for a crime. This will normally be the case when there is a formal investigation under national law and a certain (official) procedure (of search) for the specific suspected person (who is known) is being initiated. The question is: Where is the suspect? Location of the person in time and place is an essential component of targeted searches carried out in the framework of investigations. The targeted search for a person in the context of Article 26(10) AI Act refers to the use of RBI systems to track an individual. For example, in the course of an investigation the law enforcement authority has considered that Person B is the likely offender of a crime that has taken place. A post-RBI system is being used to analyse CCTV footage and video material that was collected around the crime scene (i.e. link to the location and time of the offence) to find out where the suspect came from or went to in the view of taking possible law enforcement actions against that person with possible consequences on the rights and freedoms of the person concerned . A template with the biometric data containing a single picture or audio recording of the suspect could be considered a ‘reference database’ in this context and would be compared to the collected CCTV and video material. For example , the prior authorisation by a competent judicial authority or an administrative authority pursuant to Article 26(10) AI Act would be required for:

• Repeated post-RBI queries to map movements of an identified suspect. Authorities repeatedly run post-RBI across multiple reference databases to reconstruct the movement of a suspect whose identity is already known in view of arresting that person; location in time and space of the person , even in the past, will help authorities to establish roadblocks at kilometric distance compatible with the time elapsed from the moment when the person appeared on the footage.
• Evidentiary confirmation for prosecution. Post-RBI is used specifically to produce confirmation that a known suspect appears on the crime-scene at a certain moment in a footage for evidentiary use.

Targeted search includes finding out the loca tion of concrete suspects. ‘Localisation ’ is not a defined term under the AI Act. Localisation that is not based on biometric data is outside the scope of the AI Act’s rules on biometrics, for example when following a person with a red scarf.

Law enforcement authorities frequently use AI systems for the analysis of a crime scene. In such cases, the AI system is used as a screening tool for analysing large amounts of footage material for the purpose, for example, to investigate the course of events. All kinds of material are collected from a crime scene including CCTV material, etc. The objective is to gain investigative insights into what has exactly happened by analysing appearances of persons and objects. Persons are not matched against external biometric databases, and only material that has a direct link to the crime scene is being used. The use of the system is focused on understanding the movements of the persons present around the crime scene. This is a work-step prior to initial identification, and even if there might be comparison of faces or other biometric features involved, in view of the closed dataset, this would not be considered RBI and high-risk. The same logic can be applied for AI systems used for screening child sexual abuse material (CSAM) that do not screen biometric data of individuals for identification purposes.

Localisation that is based on biometric data is considered falling under the definition of biometric categorisation when persons are not being searched on the basis of biometric templates that allow their identification, but for example based on their eye or hair colour or their height (e.g. a search for a suspect described by witnesses as a skinny woman, mid-thirties, 160-165 cm, light skin, dark hair, brown eyes, leaving the surrounding of the crime scene). The AI system issorting and filtering footage to flag persons that fall within this category. Pure searching for occurrences based on biometric characteristics – without establishing biometric templates that allows the identification is not considered RBI.

The objective of requiring authorisation for the use of a post-RBI system for the targeted search of a person suspected or convicted of having committed a criminal offence is the need for an assessment and a decision as to whether the envisaged use of the AI system for the purpose is strictly necessary for finding the person in the context of the investigation of a specific criminal offence. This necessity assessment must take account of the personal, temporal and geographic scope of the targeted search and assess whether t here are not less intrusive means for finding the person in question. The condition of strict necessity is assessed for each concrete case by the law enforcement authority intending to deploy the system. The authorisation request can be linked to the opening of an investigation or the request for other surveillance measures according to national law to streamline procedures.

According to Article 26(10), subparagraph 1, AI Act, the request must be made ex ante or without undue delay and no later than 48 hours after starting the use of the post-RBI system.

The authorisation must be requested from a judicial authority or an administrative authority whose decision is binding and subject to judicial review. The obligation to request an authorisation solely applies to the targeted search of a person suspected or convicted of having committed a criminal offence in the framework of an investigation. This excludes, for example, the targeted search for missing persons or victims (no authorisation is needed) or se arches outside the framework of an investigation of a crime (e.g., law enforcement use a biometric template of a known child sexual abuse victim to identify whether images/videos featuring the same child come up in other unrelated reports of child sexual a buse or to search for the appearances of a picture of a witness of a crime scene in an indexed version of social media , e.g. scraped material, on the basis of the witness driver’s license mugshot. Depending on national law, ) the authoritsation granted by the judicial authority or administrative authority may concern a particular person or a specific group of persons suspected of having committed or convicted of concrete criminal offence(s) in the context of a specific investigation. It may not be necessary to know the civil identity of the person(s).

The authorisation procedure required by Article 26(1 0) AI Act can be designed with sufficient flexibility to also cover unforeseeable operational scenarios and the screening of large data volumes, for example CSAM. In case of CSAM the search for victims would not be subject to an authorisation requirement. Neither would the sorting and filtering of data be relevant. The authorisation requirement would apply only in the case of a specifically targeted search against a concrete suspect (unless it is an initial identification).Without prejudice to the independence of the judiciary and to national procedural law and practices, and to facili tate the implementation of this requirement and make it operational, Member States could consider relying on the judicial authorisation procedures that are currently established for compliance with other Union or national laws, e.g., concerning an authorisation for the interception of communications.

If the requested authorisation is rejected, the use of the post-RBI system shall be stopped with immediate effect and the personal data linked to the use of the high-risk AI system for which the authorisation was requested must be deleted. This does not imply that a standard reference database is to be deleted, butonly personal data that is specific to the investigation. This requirement applies to the results of the processing of CCTV or video material , such as biometric profiles and identifications or links made through comparison of the material with reference data. It does not apply, for example, to the CCTV material itself, where this material has been lawfully obtained.

In no case shall post-RBI be used for law enforcement purposes in an untargeted way, without any link of a specific person to a criminal offence, a criminal proceeding, a genuine and present or genuine and foreseeable threat of a criminal offence, or the search for a specific missing person (Subparagraph 3). That applies universally, also to cases of post RBI use of systems for law enforcement purposes when there is no requirement for an authorisation for the targeted search of persons in the framework of an investigation as set out under Article 26(10), subparagraph 1, AI Act.

Recital 95 AI Act further specifies that ’[p]ost-RBI systems should always be used in a way that is proportionate, legitimate and strictly necessary, and thus targeted, in terms of the individuals to be identified, the location, temporal scope and based on a closed data set of legally acquired video footage. In any case, post-RBI systems should not be used in the framework of law enforcement to lead to indiscriminate surveillance. The conditions for post-RBI should in any case not provide a basis to circumvent the conditions of the prohibition and strict exceptions for real time remote biometric identification.’ Use cases are different and depend on the concrete situation. Compliance with the condition of strict necessity for the investigation of a specific criminal offence must be assessed and demonstrated by the law enforcement authority intending to deploy the system. This can also apply to intelligence operations, if they are carried out for the purpose of law enforcement and are not exclusively for national security purposes. Examples of untargeted searches include:

• Police run facial recognition across all persons appearing in city-center CCTV over a weekend, without searching for a specific person in mind, to detect possible offenders.
• Police deploy a post-RBI system after a public demonstration happened to scan all persons appearing in footage from a public square where the demonstration took place, without having previously identified a specific suspect or concrete individual linked to an offence.
• Authorities regularly check public space (e.g. transport hubs) video feeds using facial recognition to detect any persons of potential interest, without a time-bound or crime-linked investigative purpose.

Article 26(10), subparagraph 3, AI Act provides that no decision that produces an adverse legal effect on a person may be taken by the law enforcement authorities based solely on the output of a post-RBI system. This restriction comes on top of the requirement under Article 14 AI Act for human oversight. Further analysis to inform a decision can, for example, relate to the question whether a given person has been at a different place or also whether there are other reasons why the person cannot be the subject of a search²².

Article 26(10), subparagraph 4, AI Act confirms that the requirements in Article 26(10) AI Act are without prejudice to Article 9 of Regulation (EU) 2016/679 and Article 10 of Directive (EU) 2016/680 for the processing of biometric data.

According to Article 26(10), subparagraphs 5 and 6, AI Act, deployers of post-RBI systems must document each use of such systems in the relevant police file and make the documentationavailable upon request. Furthermore, they must proactively submitannual reports (that may be aggregated to cover more than one deployment) to the relevant market surveillance authority and the national data protection authority, excluding the disclosure of sensitive operational data related to law enforcement.

Pursuant to Article 26(10), subparagraph 7, AI Act, Member States may introduce, in accordance with Union law, more restrictive laws on the use of post-RBI systems.